drivers/md/dm-log-writes.c

   1 /*
   2  * Copyright (C) 2014 Facebook. All rights reserved.
   3  *
   4  * This file is released under the GPL.
   5  */
   6
   7 #include <linux/device-mapper.h>
   8
   9 #include <linux/module.h>
  10 #include <linux/init.h>
  11 #include <linux/blkdev.h>
  12 #include <linux/bio.h>
  13 #include <linux/dax.h>
  14 #include <linux/slab.h>
  15 #include <linux/kthread.h>
  16 #include <linux/freezer.h>
  17 #include <linux/uio.h>
  18
  19 #define DM_MSG_PREFIX "log-writes"
  20
  21 /*
  22  * This target will sequentially log all writes to the target device onto the
  23  * log device.  This is helpful for replaying writes to check for fs consistency
  24  * at all times.  This target provides a mechanism to mark specific events to
  25  * check data at a later time.  So for example you would:
  26  *
  27  * write data
  28  * fsync
  29  * dmsetup message /dev/whatever mark mymark
  30  * unmount /mnt/test
  31  *
  32  * Then replay the log up to mymark and check the contents of the replay to
  33  * verify it matches what was written.
  34  *
  35  * We log writes only after they have been flushed, this makes the log describe
  36  * close to the order in which the data hits the actual disk, not its cache.  So
  37  * for example the following sequence (W means write, C means complete)
  38  *
  39  * Wa,Wb,Wc,Cc,Ca,FLUSH,FUAd,Cb,CFLUSH,CFUAd
  40  *
  41  * Would result in the log looking like this:
  42  *
  43  * c,a,flush,fuad,b,<other writes>,<next flush>
  44  *
  45  * This is meant to help expose problems where file systems do not properly wait
  46  * on data being written before invoking a FLUSH.  FUA bypasses cache so once it
  47  * completes it is added to the log as it should be on disk.
  48  *
  49  * We treat DISCARDs as if they don't bypass cache so that they are logged in
  50  * order of completion along with the normal writes.  If we didn't do it this
  51  * way we would process all the discards first and then write all the data, when
  52  * in fact we want to do the data and the discard in the order that they
  53  * completed.
  54  */
  55 #define LOG_FLUSH_FLAG          (1 << 0)
  56 #define LOG_FUA_FLAG            (1 << 1)
  57 #define LOG_DISCARD_FLAG        (1 << 2)
  58 #define LOG_MARK_FLAG           (1 << 3)
  59 #define LOG_METADATA_FLAG       (1 << 4)
  60
  61 #define WRITE_LOG_VERSION 1ULL
  62 #define WRITE_LOG_MAGIC 0x6a736677736872ULL
  63
  64 /*
  65  * The disk format for this is braindead simple.
  66  *
  67  * At byte 0 we have our super, followed by the following sequence for
  68  * nr_entries:
  69  *
  70  * [   1 sector    ][  entry->nr_sectors ]
  71  * [log_write_entry][    data written    ]
  72  *
  73  * The log_write_entry takes up a full sector so we can have arbitrary length
  74  * marks and it leaves us room for extra content in the future.
  75  */
  76
  77 /*
  78  * Basic info about the log for userspace.
  79  */
  80 struct log_write_super {
  81         __le64 magic;
  82         __le64 version;
  83         __le64 nr_entries;
  84         __le32 sectorsize;
  85 };
  86
  87 /*
  88  * sector - the sector we wrote.
  89  * nr_sectors - the number of sectors we wrote.
  90  * flags - flags for this log entry.
  91  * data_len - the size of the data in this log entry, this is for private log
  92  * entry stuff, the MARK data provided by userspace for example.
  93  */
  94 struct log_write_entry {
  95         __le64 sector;
  96         __le64 nr_sectors;
  97         __le64 flags;
  98         __le64 data_len;
  99 };
 100
 101 struct log_writes_c {
 102         struct dm_dev *dev;
 103         struct dm_dev *logdev;
 104         u64 logged_entries;
 105         u32 sectorsize;
 106         u32 sectorshift;
 107         atomic_t io_blocks;
 108         atomic_t pending_blocks;
 109         sector_t next_sector;
 110         sector_t end_sector;
 111         bool logging_enabled;
 112         bool device_supports_discard;
 113         spinlock_t blocks_lock;
 114         struct list_head unflushed_blocks;
 115         struct list_head logging_blocks;
 116         wait_queue_head_t wait;
 117         struct task_struct *log_kthread;
 118 };
 119
 120 struct pending_block {
 121         int vec_cnt;
 122         u64 flags;
 123         sector_t sector;
 124         sector_t nr_sectors;
 125         char *data;
 126         u32 datalen;
 127         struct list_head list;
 128         struct bio_vec vecs[0];
 129 };
 130
 131 struct per_bio_data {
 132         struct pending_block *block;
 133 };
 134
 135 static inline sector_t bio_to_dev_sectors(struct log_writes_c *lc,
 136                                           sector_t sectors)
 137 {
 138         return sectors >> (lc->sectorshift - SECTOR_SHIFT);
 139 }
 140
 141 static inline sector_t dev_to_bio_sectors(struct log_writes_c *lc,
 142                                           sector_t sectors)
 143 {
 144         return sectors << (lc->sectorshift - SECTOR_SHIFT);
 145 }
 146
 147 static void put_pending_block(struct log_writes_c *lc)
 148 {
 149         if (atomic_dec_and_test(&lc->pending_blocks)) {
 150                 smp_mb__after_atomic();
 151                 if (waitqueue_active(&lc->wait))
 152                         wake_up(&lc->wait);
 153         }
 154 }
 155
 156 static void put_io_block(struct log_writes_c *lc)
 157 {
 158         if (atomic_dec_and_test(&lc->io_blocks)) {
 159                 smp_mb__after_atomic();
 160                 if (waitqueue_active(&lc->wait))
 161                         wake_up(&lc->wait);
 162         }
 163 }
 164
 165 static void log_end_io(struct bio *bio)
 166 {
 167         struct log_writes_c *lc = bio->bi_private;
 168
 169         if (bio->bi_status) {
 170                 unsigned long flags;
 171
 172                 DMERR("Error writing log block, error=%d", bio->bi_status);
 173                 spin_lock_irqsave(&lc->blocks_lock, flags);
 174                 lc->logging_enabled = false;
 175                 spin_unlock_irqrestore(&lc->blocks_lock, flags);
 176         }
 177
 178         bio_free_pages(bio);
 179         put_io_block(lc);
 180         bio_put(bio);
 181 }
 182
 183 /*
 184  * Meant to be called if there is an error, it will free all the pages
 185  * associated with the block.
 186  */
 187 static void free_pending_block(struct log_writes_c *lc,
 188                                struct pending_block *block)
 189 {
 190         int i;
 191
 192         for (i = 0; i < block->vec_cnt; i++) {
 193                 if (block->vecs[i].bv_page)
 194                         __free_page(block->vecs[i].bv_page);
 195         }
 196         kfree(block->data);
 197         kfree(block);
 198         put_pending_block(lc);
 199 }
 200
 201 static int write_metadata(struct log_writes_c *lc, void *entry,
 202                           size_t entrylen, void *data, size_t datalen,
 203                           sector_t sector)
 204 {
 205         struct bio *bio;
 206         struct page *page;
 207         void *ptr;
 208         size_t ret;
 209
 210         bio = bio_alloc(GFP_KERNEL, 1);
 211         if (!bio) {
 212                 DMERR("Couldn't alloc log bio");
 213                 goto error;
 214         }
 215         bio->bi_iter.bi_size = 0;
 216         bio->bi_iter.bi_sector = sector;
 217         bio_set_dev(bio, lc->logdev->bdev);
 218         bio->bi_end_io = log_end_io;
 219         bio->bi_private = lc;
 220         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 221
 222         page = alloc_page(GFP_KERNEL);
 223         if (!page) {
 224                 DMERR("Couldn't alloc log page");
 225                 bio_put(bio);
 226                 goto error;
 227         }
 228
 229         ptr = kmap_atomic(page);
 230         memcpy(ptr, entry, entrylen);
 231         if (datalen)
 232                 memcpy(ptr + entrylen, data, datalen);
 233         memset(ptr + entrylen + datalen, 0,
 234                lc->sectorsize - entrylen - datalen);
 235         kunmap_atomic(ptr);
 236
 237         ret = bio_add_page(bio, page, lc->sectorsize, 0);
 238         if (ret != lc->sectorsize) {
 239                 DMERR("Couldn't add page to the log block");
 240                 goto error_bio;
 241         }
 242         submit_bio(bio);
 243         return 0;
 244 error_bio:
 245         bio_put(bio);
 246         __free_page(page);
 247 error:
 248         put_io_block(lc);
 249         return -1;
 250 }
 251
 252 static int write_inline_data(struct log_writes_c *lc, void *entry,
 253                              size_t entrylen, void *data, size_t datalen,
 254                              sector_t sector)
 255 {
 256         int num_pages, bio_pages, pg_datalen, pg_sectorlen, i;
 257         struct page *page;
 258         struct bio *bio;
 259         size_t ret;
 260         void *ptr;
 261
 262         while (datalen) {
 263                 num_pages = ALIGN(datalen, PAGE_SIZE) >> PAGE_SHIFT;
 264                 bio_pages = min(num_pages, BIO_MAX_PAGES);
 265
 266                 atomic_inc(&lc->io_blocks);
 267
 268                 bio = bio_alloc(GFP_KERNEL, bio_pages);
 269                 if (!bio) {
 270                         DMERR("Couldn't alloc inline data bio");
 271                         goto error;
 272                 }
 273
 274                 bio->bi_iter.bi_size = 0;
 275                 bio->bi_iter.bi_sector = sector;
 276                 bio_set_dev(bio, lc->logdev->bdev);
 277                 bio->bi_end_io = log_end_io;
 278                 bio->bi_private = lc;
 279                 bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 280
 281                 for (i = 0; i < bio_pages; i++) {
 282                         pg_datalen = min_t(int, datalen, PAGE_SIZE);
 283                         pg_sectorlen = ALIGN(pg_datalen, lc->sectorsize);
 284
 285                         page = alloc_page(GFP_KERNEL);
 286                         if (!page) {
 287                                 DMERR("Couldn't alloc inline data page");
 288                                 goto error_bio;
 289                         }
 290
 291                         ptr = kmap_atomic(page);
 292                         memcpy(ptr, data, pg_datalen);
 293                         if (pg_sectorlen > pg_datalen)
 294                                 memset(ptr + pg_datalen, 0, pg_sectorlen - pg_datalen);
 295                         kunmap_atomic(ptr);
 296
 297                         ret = bio_add_page(bio, page, pg_sectorlen, 0);
 298                         if (ret != pg_sectorlen) {
 299                                 DMERR("Couldn't add page of inline data");
 300                                 __free_page(page);
 301                                 goto error_bio;
 302                         }
 303
 304                         datalen -= pg_datalen;
 305                         data    += pg_datalen;
 306                 }
 307                 submit_bio(bio);
 308
 309                 sector += bio_pages * PAGE_SECTORS;
 310         }
 311         return 0;
 312 error_bio:
 313         bio_free_pages(bio);
 314         bio_put(bio);
 315 error:
 316         put_io_block(lc);
 317         return -1;
 318 }
 319
 320 static int log_one_block(struct log_writes_c *lc,
 321                          struct pending_block *block, sector_t sector)
 322 {
 323         struct bio *bio;
 324         struct log_write_entry entry;
 325         size_t metadatalen, ret;
 326         int i;
 327
 328         entry.sector = cpu_to_le64(block->sector);
 329         entry.nr_sectors = cpu_to_le64(block->nr_sectors);
 330         entry.flags = cpu_to_le64(block->flags);
 331         entry.data_len = cpu_to_le64(block->datalen);
 332
 333         metadatalen = (block->flags & LOG_MARK_FLAG) ? block->datalen : 0;
 334         if (write_metadata(lc, &entry, sizeof(entry), block->data,
 335                            metadatalen, sector)) {
 336                 free_pending_block(lc, block);
 337                 return -1;
 338         }
 339
 340         sector += dev_to_bio_sectors(lc, 1);
 341
 342         if (block->datalen && metadatalen == 0) {
 343                 if (write_inline_data(lc, &entry, sizeof(entry), block->data,
 344                                       block->datalen, sector)) {
 345                         free_pending_block(lc, block);
 346                         return -1;
 347                 }
 348                 /* we don't support both inline data & bio data */
 349                 goto out;
 350         }
 351
 352         if (!block->vec_cnt)
 353                 goto out;
 354
 355         atomic_inc(&lc->io_blocks);
 356         bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt, BIO_MAX_PAGES));
 357         if (!bio) {
 358                 DMERR("Couldn't alloc log bio");
 359                 goto error;
 360         }
 361         bio->bi_iter.bi_size = 0;
 362         bio->bi_iter.bi_sector = sector;
 363         bio_set_dev(bio, lc->logdev->bdev);
 364         bio->bi_end_io = log_end_io;
 365         bio->bi_private = lc;
 366         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 367
 368         for (i = 0; i < block->vec_cnt; i++) {
 369                 /*
 370                  * The page offset is always 0 because we allocate a new page
 371                  * for every bvec in the original bio for simplicity sake.
 372                  */
 373                 ret = bio_add_page(bio, block->vecs[i].bv_page,
 374                                    block->vecs[i].bv_len, 0);
 375                 if (ret != block->vecs[i].bv_len) {
 376                         atomic_inc(&lc->io_blocks);
 377                         submit_bio(bio);
 378                         bio = bio_alloc(GFP_KERNEL, min(block->vec_cnt - i, BIO_MAX_PAGES));
 379                         if (!bio) {
 380                                 DMERR("Couldn't alloc log bio");
 381                                 goto error;
 382                         }
 383                         bio->bi_iter.bi_size = 0;
 384                         bio->bi_iter.bi_sector = sector;
 385                         bio_set_dev(bio, lc->logdev->bdev);
 386                         bio->bi_end_io = log_end_io;
 387                         bio->bi_private = lc;
 388                         bio_set_op_attrs(bio, REQ_OP_WRITE, 0);
 389
 390                         ret = bio_add_page(bio, block->vecs[i].bv_page,
 391                                            block->vecs[i].bv_len, 0);
 392                         if (ret != block->vecs[i].bv_len) {
 393                                 DMERR("Couldn't add page on new bio?");
 394                                 bio_put(bio);
 395                                 goto error;
 396                         }
 397                 }
 398                 sector += block->vecs[i].bv_len >> SECTOR_SHIFT;
 399         }
 400         submit_bio(bio);
 401 out:
 402         kfree(block->data);
 403         kfree(block);
 404         put_pending_block(lc);
 405         return 0;
 406 error:
 407         free_pending_block(lc, block);
 408         put_io_block(lc);
 409         return -1;
 410 }
 411
 412 static int log_super(struct log_writes_c *lc)
 413 {
 414         struct log_write_super super;
 415
 416         super.magic = cpu_to_le64(WRITE_LOG_MAGIC);
 417         super.version = cpu_to_le64(WRITE_LOG_VERSION);
 418         super.nr_entries = cpu_to_le64(lc->logged_entries);
 419         super.sectorsize = cpu_to_le32(lc->sectorsize);
 420
 421         if (write_metadata(lc, &super, sizeof(super), NULL, 0, 0)) {
 422                 DMERR("Couldn't write super");
 423                 return -1;
 424         }
 425
 426         return 0;
 427 }
 428
 429 static inline sector_t logdev_last_sector(struct log_writes_c *lc)
 430 {
 431         return i_size_read(lc->logdev->bdev->bd_inode) >> SECTOR_SHIFT;
 432 }
 433
 434 static int log_writes_kthread(void *arg)
 435 {
 436         struct log_writes_c *lc = (struct log_writes_c *)arg;
 437         sector_t sector = 0;
 438
 439         while (!kthread_should_stop()) {
 440                 bool super = false;
 441                 bool logging_enabled;
 442                 struct pending_block *block = NULL;
 443                 int ret;
 444
 445                 spin_lock_irq(&lc->blocks_lock);
 446                 if (!list_empty(&lc->logging_blocks)) {
 447                         block = list_first_entry(&lc->logging_blocks,
 448                                                  struct pending_block, list);
 449                         list_del_init(&block->list);
 450                         if (!lc->logging_enabled)
 451                                 goto next;
 452
 453                         sector = lc->next_sector;
 454                         if (!(block->flags & LOG_DISCARD_FLAG))
 455                                 lc->next_sector += dev_to_bio_sectors(lc, block->nr_sectors);
 456                         lc->next_sector += dev_to_bio_sectors(lc, 1);
 457
 458                         /*
 459                          * Apparently the size of the device may not be known
 460                          * right away, so handle this properly.
 461                          */
 462                         if (!lc->end_sector)
 463                                 lc->end_sector = logdev_last_sector(lc);
 464                         if (lc->end_sector &&
 465                             lc->next_sector >= lc->end_sector) {
 466                                 DMERR("Ran out of space on the logdev");
 467                                 lc->logging_enabled = false;
 468                                 goto next;
 469                         }
 470                         lc->logged_entries++;
 471                         atomic_inc(&lc->io_blocks);
 472
 473                         super = (block->flags & (LOG_FUA_FLAG | LOG_MARK_FLAG));
 474                         if (super)
 475                                 atomic_inc(&lc->io_blocks);
 476                 }
 477 next:
 478                 logging_enabled = lc->logging_enabled;
 479                 spin_unlock_irq(&lc->blocks_lock);
 480                 if (block) {
 481                         if (logging_enabled) {
 482                                 ret = log_one_block(lc, block, sector);
 483                                 if (!ret && super)
 484                                         ret = log_super(lc);
 485                                 if (ret) {
 486                                         spin_lock_irq(&lc->blocks_lock);
 487                                         lc->logging_enabled = false;
 488                                         spin_unlock_irq(&lc->blocks_lock);
 489                                 }
 490                         } else
 491                                 free_pending_block(lc, block);
 492                         continue;
 493                 }
 494
 495                 if (!try_to_freeze()) {
 496                         set_current_state(TASK_INTERRUPTIBLE);
 497                         if (!kthread_should_stop() &&
 498                             list_empty(&lc->logging_blocks))
 499                                 schedule();
 500                         __set_current_state(TASK_RUNNING);
 501                 }
 502         }
 503         return 0;
 504 }
 505
 506 /*
 507  * Construct a log-writes mapping:
 508  * log-writes <dev_path> <log_dev_path>
 509  */
 510 static int log_writes_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 511 {
 512         struct log_writes_c *lc;
 513         struct dm_arg_set as;
 514         const char *devname, *logdevname;
 515         int ret;
 516
 517         as.argc = argc;
 518         as.argv = argv;
 519
 520         if (argc < 2) {
 521                 ti->error = "Invalid argument count";
 522                 return -EINVAL;
 523         }
 524
 525         lc = kzalloc(sizeof(struct log_writes_c), GFP_KERNEL);
 526         if (!lc) {
 527                 ti->error = "Cannot allocate context";
 528                 return -ENOMEM;
 529         }
 530         spin_lock_init(&lc->blocks_lock);
 531         INIT_LIST_HEAD(&lc->unflushed_blocks);
 532         INIT_LIST_HEAD(&lc->logging_blocks);
 533         init_waitqueue_head(&lc->wait);
 534         atomic_set(&lc->io_blocks, 0);
 535         atomic_set(&lc->pending_blocks, 0);
 536
 537         devname = dm_shift_arg(&as);
 538         ret = dm_get_device(ti, devname, dm_table_get_mode(ti->table), &lc->dev);
 539         if (ret) {
 540                 ti->error = "Device lookup failed";
 541                 goto bad;
 542         }
 543
 544         logdevname = dm_shift_arg(&as);
 545         ret = dm_get_device(ti, logdevname, dm_table_get_mode(ti->table),
 546                             &lc->logdev);
 547         if (ret) {
 548                 ti->error = "Log device lookup failed";
 549                 dm_put_device(ti, lc->dev);
 550                 goto bad;
 551         }
 552
 553         lc->sectorsize = bdev_logical_block_size(lc->dev->bdev);
 554         lc->sectorshift = ilog2(lc->sectorsize);
 555         lc->log_kthread = kthread_run(log_writes_kthread, lc, "log-write");
 556         if (IS_ERR(lc->log_kthread)) {
 557                 ret = PTR_ERR(lc->log_kthread);
 558                 ti->error = "Couldn't alloc kthread";
 559                 dm_put_device(ti, lc->dev);
 560                 dm_put_device(ti, lc->logdev);
 561                 goto bad;
 562         }
 563
 564         /*
 565          * next_sector is in 512b sectors to correspond to what bi_sector expects.
 566          * The super starts at sector 0, and the next_sector is the next logical
 567          * one based on the sectorsize of the device.
 568          */
 569         lc->next_sector = lc->sectorsize >> SECTOR_SHIFT;
 570         lc->logging_enabled = true;
 571         lc->end_sector = logdev_last_sector(lc);
 572         lc->device_supports_discard = true;
 573
 574         ti->num_flush_bios = 1;
 575         ti->flush_supported = true;
 576         ti->num_discard_bios = 1;
 577         ti->discards_supported = true;
 578         ti->per_io_data_size = sizeof(struct per_bio_data);
 579         ti->private = lc;
 580         return 0;
 581
 582 bad:
 583         kfree(lc);
 584         return ret;
 585 }
 586
 587 static int log_mark(struct log_writes_c *lc, char *data)
 588 {
 589         struct pending_block *block;
 590         size_t maxsize = lc->sectorsize - sizeof(struct log_write_entry);
 591
 592         block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
 593         if (!block) {
 594                 DMERR("Error allocating pending block");
 595                 return -ENOMEM;
 596         }
 597
 598         block->data = kstrndup(data, maxsize - 1, GFP_KERNEL);
 599         if (!block->data) {
 600                 DMERR("Error copying mark data");
 601                 kfree(block);
 602                 return -ENOMEM;
 603         }
 604         atomic_inc(&lc->pending_blocks);
 605         block->datalen = strlen(block->data);
 606         block->flags |= LOG_MARK_FLAG;
 607         spin_lock_irq(&lc->blocks_lock);
 608         list_add_tail(&block->list, &lc->logging_blocks);
 609         spin_unlock_irq(&lc->blocks_lock);
 610         wake_up_process(lc->log_kthread);
 611         return 0;
 612 }
 613
 614 static int log_dax(struct log_writes_c *lc, sector_t sector, size_t bytes,
 615                    struct iov_iter *i)
 616 {
 617         struct pending_block *block;
 618
 619         if (!bytes)
 620                 return 0;
 621
 622         block = kzalloc(sizeof(struct pending_block), GFP_KERNEL);
 623         if (!block) {
 624                 DMERR("Error allocating dax pending block");
 625                 return -ENOMEM;
 626         }
 627
 628         block->data = kzalloc(bytes, GFP_KERNEL);
 629         if (!block->data) {
 630                 DMERR("Error allocating dax data space");
 631                 kfree(block);
 632                 return -ENOMEM;
 633         }
 634
 635         /* write data provided via the iterator */
 636         if (!copy_from_iter(block->data, bytes, i)) {
 637                 DMERR("Error copying dax data");
 638                 kfree(block->data);
 639                 kfree(block);
 640                 return -EIO;
 641         }
 642
 643         /* rewind the iterator so that the block driver can use it */
 644         iov_iter_revert(i, bytes);
 645
 646         block->datalen = bytes;
 647         block->sector = bio_to_dev_sectors(lc, sector);
 648         block->nr_sectors = ALIGN(bytes, lc->sectorsize) >> lc->sectorshift;
 649
 650         atomic_inc(&lc->pending_blocks);
 651         spin_lock_irq(&lc->blocks_lock);
 652         list_add_tail(&block->list, &lc->unflushed_blocks);
 653         spin_unlock_irq(&lc->blocks_lock);
 654         wake_up_process(lc->log_kthread);
 655
 656         return 0;
 657 }
 658
 659 static void log_writes_dtr(struct dm_target *ti)
 660 {
 661         struct log_writes_c *lc = ti->private;
 662
 663         spin_lock_irq(&lc->blocks_lock);
 664         list_splice_init(&lc->unflushed_blocks, &lc->logging_blocks);
 665         spin_unlock_irq(&lc->blocks_lock);
 666
 667         /*
 668          * This is just nice to have since it'll update the super to include the
 669          * unflushed blocks, if it fails we don't really care.
 670          */
 671         log_mark(lc, "dm-log-writes-end");
 672         wake_up_process(lc->log_kthread);
 673         wait_event(lc->wait, !atomic_read(&lc->io_blocks) &&
 674                    !atomic_read(&lc->pending_blocks));
 675         kthread_stop(lc->log_kthread);
 676
 677         WARN_ON(!list_empty(&lc->logging_blocks));
 678         WARN_ON(!list_empty(&lc->unflushed_blocks));
 679         dm_put_device(ti, lc->dev);
 680         dm_put_device(ti, lc->logdev);
 681         kfree(lc);
 682 }
 683
 684 static void normal_map_bio(struct dm_target *ti, struct bio *bio)
 685 {
 686         struct log_writes_c *lc = ti->private;
 687
 688         bio_set_dev(bio, lc->dev->bdev);
 689 }
 690
 691 static int log_writes_map(struct dm_target *ti, struct bio *bio)
 692 {
 693         struct log_writes_c *lc = ti->private;
 694         struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 695         struct pending_block *block;
 696         struct bvec_iter iter;
 697         struct bio_vec bv;
 698         size_t alloc_size;
 699         int i = 0;
 700         bool flush_bio = (bio->bi_opf & REQ_PREFLUSH);
 701         bool fua_bio = (bio->bi_opf & REQ_FUA);
 702         bool discard_bio = (bio_op(bio) == REQ_OP_DISCARD);
 703         bool meta_bio = (bio->bi_opf & REQ_META);
 704
 705         pb->block = NULL;
 706
 707         /* Don't bother doing anything if logging has been disabled */
 708         if (!lc->logging_enabled)
 709                 goto map_bio;
 710
 711         /*
 712          * Map reads as normal.
 713          */
 714         if (bio_data_dir(bio) == READ)
 715                 goto map_bio;
 716
 717         /* No sectors and not a flush?  Don't care */
 718         if (!bio_sectors(bio) && !flush_bio)
 719                 goto map_bio;
 720
 721         /*
 722          * Discards will have bi_size set but there's no actual data, so just
 723          * allocate the size of the pending block.
 724          */
 725         if (discard_bio)
 726                 alloc_size = sizeof(struct pending_block);
 727         else
 728                 alloc_size = sizeof(struct pending_block) + sizeof(struct bio_vec) * bio_segments(bio);
 729
 730         block = kzalloc(alloc_size, GFP_NOIO);
 731         if (!block) {
 732                 DMERR("Error allocating pending block");
 733                 spin_lock_irq(&lc->blocks_lock);
 734                 lc->logging_enabled = false;
 735                 spin_unlock_irq(&lc->blocks_lock);
 736                 return DM_MAPIO_KILL;
 737         }
 738         INIT_LIST_HEAD(&block->list);
 739         pb->block = block;
 740         atomic_inc(&lc->pending_blocks);
 741
 742         if (flush_bio)
 743                 block->flags |= LOG_FLUSH_FLAG;
 744         if (fua_bio)
 745                 block->flags |= LOG_FUA_FLAG;
 746         if (discard_bio)
 747                 block->flags |= LOG_DISCARD_FLAG;
 748         if (meta_bio)
 749                 block->flags |= LOG_METADATA_FLAG;
 750
 751         block->sector = bio_to_dev_sectors(lc, bio->bi_iter.bi_sector);
 752         block->nr_sectors = bio_to_dev_sectors(lc, bio_sectors(bio));
 753
 754         /* We don't need the data, just submit */
 755         if (discard_bio) {
 756                 WARN_ON(flush_bio || fua_bio);
 757                 if (lc->device_supports_discard)
 758                         goto map_bio;
 759                 bio_endio(bio);
 760                 return DM_MAPIO_SUBMITTED;
 761         }
 762
 763         /* Flush bio, splice the unflushed blocks onto this list and submit */
 764         if (flush_bio && !bio_sectors(bio)) {
 765                 spin_lock_irq(&lc->blocks_lock);
 766                 list_splice_init(&lc->unflushed_blocks, &block->list);
 767                 spin_unlock_irq(&lc->blocks_lock);
 768                 goto map_bio;
 769         }
 770
 771         /*
 772          * We will write this bio somewhere else way later so we need to copy
 773          * the actual contents into new pages so we know the data will always be
 774          * there.
 775          *
 776          * We do this because this could be a bio from O_DIRECT in which case we
 777          * can't just hold onto the page until some later point, we have to
 778          * manually copy the contents.
 779          */
 780         bio_for_each_segment(bv, bio, iter) {
 781                 struct page *page;
 782                 void *src, *dst;
 783
 784                 page = alloc_page(GFP_NOIO);
 785                 if (!page) {
 786                         DMERR("Error allocing page");
 787                         free_pending_block(lc, block);
 788                         spin_lock_irq(&lc->blocks_lock);
 789                         lc->logging_enabled = false;
 790                         spin_unlock_irq(&lc->blocks_lock);
 791                         return DM_MAPIO_KILL;
 792                 }
 793
 794                 src = kmap_atomic(bv.bv_page);
 795                 dst = kmap_atomic(page);
 796                 memcpy(dst, src + bv.bv_offset, bv.bv_len);
 797                 kunmap_atomic(dst);
 798                 kunmap_atomic(src);
 799                 block->vecs[i].bv_page = page;
 800                 block->vecs[i].bv_len = bv.bv_len;
 801                 block->vec_cnt++;
 802                 i++;
 803         }
 804
 805         /* Had a flush with data in it, weird */
 806         if (flush_bio) {
 807                 spin_lock_irq(&lc->blocks_lock);
 808                 list_splice_init(&lc->unflushed_blocks, &block->list);
 809                 spin_unlock_irq(&lc->blocks_lock);
 810         }
 811 map_bio:
 812         normal_map_bio(ti, bio);
 813         return DM_MAPIO_REMAPPED;
 814 }
 815
 816 static int normal_end_io(struct dm_target *ti, struct bio *bio,
 817                 blk_status_t *error)
 818 {
 819         struct log_writes_c *lc = ti->private;
 820         struct per_bio_data *pb = dm_per_bio_data(bio, sizeof(struct per_bio_data));
 821
 822         if (bio_data_dir(bio) == WRITE && pb->block) {
 823                 struct pending_block *block = pb->block;
 824                 unsigned long flags;
 825
 826                 spin_lock_irqsave(&lc->blocks_lock, flags);
 827                 if (block->flags & LOG_FLUSH_FLAG) {
 828                         list_splice_tail_init(&block->list, &lc->logging_blocks);
 829                         list_add_tail(&block->list, &lc->logging_blocks);
 830                         wake_up_process(lc->log_kthread);
 831                 } else if (block->flags & LOG_FUA_FLAG) {
 832                         list_add_tail(&block->list, &lc->logging_blocks);
 833                         wake_up_process(lc->log_kthread);
 834                 } else
 835                         list_add_tail(&block->list, &lc->unflushed_blocks);
 836                 spin_unlock_irqrestore(&lc->blocks_lock, flags);
 837         }
 838
 839         return DM_ENDIO_DONE;
 840 }
 841
 842 /*
 843  * INFO format: <logged entries> <highest allocated sector>
 844  */
 845 static void log_writes_status(struct dm_target *ti, status_type_t type,
 846                               unsigned status_flags, char *result,
 847                               unsigned maxlen)
 848 {
 849         unsigned sz = 0;
 850         struct log_writes_c *lc = ti->private;
 851
 852         switch (type) {
 853         case STATUSTYPE_INFO:
 854                 DMEMIT("%llu %llu", lc->logged_entries,
 855                        (unsigned long long)lc->next_sector - 1);
 856                 if (!lc->logging_enabled)
 857                         DMEMIT(" logging_disabled");
 858                 break;
 859
 860         case STATUSTYPE_TABLE:
 861                 DMEMIT("%s %s", lc->dev->name, lc->logdev->name);
 862                 break;
 863         }
 864 }
 865
 866 static int log_writes_prepare_ioctl(struct dm_target *ti,
 867                 struct block_device **bdev, fmode_t *mode)
 868 {
 869         struct log_writes_c *lc = ti->private;
 870         struct dm_dev *dev = lc->dev;
 871
 872         *bdev = dev->bdev;
 873         /*
 874          * Only pass ioctls through if the device sizes match exactly.
 875          */
 876         if (ti->len != i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT)
 877                 return 1;
 878         return 0;
 879 }
 880
 881 static int log_writes_iterate_devices(struct dm_target *ti,
 882                                       iterate_devices_callout_fn fn,
 883                                       void *data)
 884 {
 885         struct log_writes_c *lc = ti->private;
 886
 887         return fn(ti, lc->dev, 0, ti->len, data);
 888 }
 889
 890 /*
 891  * Messages supported:
 892  *   mark <mark data> - specify the marked data.
 893  */
 894 static int log_writes_message(struct dm_target *ti, unsigned argc, char **argv,
 895                               char *result, unsigned maxlen)
 896 {
 897         int r = -EINVAL;
 898         struct log_writes_c *lc = ti->private;
 899
 900         if (argc != 2) {
 901                 DMWARN("Invalid log-writes message arguments, expect 2 arguments, got %d", argc);
 902                 return r;
 903         }
 904
 905         if (!strcasecmp(argv[0], "mark"))
 906                 r = log_mark(lc, argv[1]);
 907         else
 908                 DMWARN("Unrecognised log writes target message received: %s", argv[0]);
 909
 910         return r;
 911 }
 912
 913 static void log_writes_io_hints(struct dm_target *ti, struct queue_limits *limits)
 914 {
 915         struct log_writes_c *lc = ti->private;
 916         struct request_queue *q = bdev_get_queue(lc->dev->bdev);
 917
 918         if (!q || !blk_queue_discard(q)) {
 919                 lc->device_supports_discard = false;
 920                 limits->discard_granularity = lc->sectorsize;
 921                 limits->max_discard_sectors = (UINT_MAX >> SECTOR_SHIFT);
 922         }
 923         limits->logical_block_size = bdev_logical_block_size(lc->dev->bdev);
 924         limits->physical_block_size = bdev_physical_block_size(lc->dev->bdev);
 925         limits->io_min = limits->physical_block_size;
 926 }
 927
 928 static long log_writes_dax_direct_access(struct dm_target *ti, pgoff_t pgoff,
 929                                          long nr_pages, void **kaddr, pfn_t *pfn)
 930 {
 931         struct log_writes_c *lc = ti->private;
 932         sector_t sector = pgoff * PAGE_SECTORS;
 933         int ret;
 934
 935         ret = bdev_dax_pgoff(lc->dev->bdev, sector, nr_pages * PAGE_SIZE, &pgoff);
 936         if (ret)
 937                 return ret;
 938         return dax_direct_access(lc->dev->dax_dev, pgoff, nr_pages, kaddr, pfn);
 939 }
 940
 941 static size_t log_writes_dax_copy_from_iter(struct dm_target *ti,
 942                                             pgoff_t pgoff, void *addr, size_t bytes,
 943                                             struct iov_iter *i)
 944 {
 945         struct log_writes_c *lc = ti->private;
 946         sector_t sector = pgoff * PAGE_SECTORS;
 947         int err;
 948
 949         if (bdev_dax_pgoff(lc->dev->bdev, sector, ALIGN(bytes, PAGE_SIZE), &pgoff))
 950                 return 0;
 951
 952         /* Don't bother doing anything if logging has been disabled */
 953         if (!lc->logging_enabled)
 954                 goto dax_copy;
 955
 956         err = log_dax(lc, sector, bytes, i);
 957         if (err) {
 958                 DMWARN("Error %d logging DAX write", err);
 959                 return 0;
 960         }
 961 dax_copy:
 962         return dax_copy_from_iter(lc->dev->dax_dev, pgoff, addr, bytes, i);
 963 }
 964
 965 static struct target_type log_writes_target = {
 966         .name   = "log-writes",
 967         .version = {1, 1, 0},
 968         .module = THIS_MODULE,
 969         .ctr    = log_writes_ctr,
 970         .dtr    = log_writes_dtr,
 971         .map    = log_writes_map,
 972         .end_io = normal_end_io,
 973         .status = log_writes_status,
 974         .prepare_ioctl = log_writes_prepare_ioctl,
 975         .message = log_writes_message,
 976         .iterate_devices = log_writes_iterate_devices,
 977         .io_hints = log_writes_io_hints,
 978         .direct_access = log_writes_dax_direct_access,
 979         .dax_copy_from_iter = log_writes_dax_copy_from_iter,
 980 };
 981
 982 static int __init dm_log_writes_init(void)
 983 {
 984         int r = dm_register_target(&log_writes_target);
 985
 986         if (r < 0)
 987                 DMERR("register failed %d", r);
 988
 989         return r;
 990 }
 991
 992 static void __exit dm_log_writes_exit(void)
 993 {
 994         dm_unregister_target(&log_writes_target);
 995 }
 996
 997 module_init(dm_log_writes_init);
 998 module_exit(dm_log_writes_exit);
 999
1000 MODULE_DESCRIPTION(DM_NAME " log writes target");
1001 MODULE_AUTHOR("Josef Bacik <jbacik@fb.com>");
1002 MODULE_LICENSE("GPL");